Resurrecting Driver Workload, Multivariate Analysis of Test Track Data

Abstract

This paper presents multivariate analyses of data collected from divers at a test track during the Driver Workload Metrics (DWM) project. As noted in a prior publication, the DWM project was a cooperative effort with the National Highway Transportation Safety Administration (NHTSA) and four automotive manufacturers. The DWM project defined workload as the competition in driver resources (perceptual, cognitive, or physical) between the driving task and a concurrent secondary task, occurring over that task’s duration. It was hypothesized that, depending on the type of secondary task performed while driving, measured workload and the correlated quality of driving should either remain the same or decline, but would manifest in degraded measures of lane keeping, longitudinal control, or eye glance behavior. Data for this new analysis was collected from test subjects who drove an instrumented car on a test track while performing various on-board tasks. These data also contain additional responses from several new visual manual task that were originally deemed to be too hazardous for test subjects while driving on a major four lane highway. It was therefore further hypothesized that the new task would demonstrate higher levels of visual-manual workload when compared to less demanding tasks. As in the prior DWM multivariate paper, test subject responses from the kinematic and eye glance behavior from the test track data were first analyzed using Maximum Likelihood Factor Analysis. This well-known statistical method attempts to uncover the underlying unobserved structure within the large set of variables. It is this hidden multi-dimensional structure that must be examined to empirically comprehend the concept of driver workload. As in the DWM on-road analyses, these new analyses found that task-induced workload affected driving performance and was multi-dimensional in nature. Visual-manual tasks exhibited fundamentally different performance profiles than auditory-vocal tasks or just driving. Furthermore, when secondary statistical analyses of the normalized factor scores were done using Multivariate Analysis of Variance (MANOVA) the results found highly statistically significant workload differences in age groups and task type.